littlelearner-5b-grpo-math-expert

littlelearner-5b-grpo-math-expert

5B K-5-bounded chat model post-trained with GRPO on top of SFT.

Part of the LittleLearner scale-up study (pedagogically-controlled knowledge exposure): Qwen3 dense LMs trained on a corpus filtered to U.S. K-5 material (bounded) vs an unfiltered FineWeb-Edu corpus (unbounded), to measure what an interpretable knowledge boundary costs and grants.

Note: This checkpoint was post-trained with GRPO on mathematical reasoning tasks to probe achievable performance on MathCAMPS. As a result, its behavior is specialized toward mathematical reasoning and may not preserve general-purpose chat capabilities; responses may also exhibit a tendency toward math-oriented reasoning or output.

Model

• Architecture: Qwen3 dense (Qwen3ForCausalLM).
• Size: 5.04B params, hidden 3072, 44 layers, 24 query / 8 KV heads, FFN 9216. Context: 4096.
• Tokenizer: custom 64k byte-level BPE with per-digit splitting (ChatML special tokens).
• Pretraining: 88B tokens on K-5 LittleCurriculum (FineWeb-Edu filtered to U.S. grades K-5). WSD schedule, sharded Muon, MXFP8, Megatron-Core on 8xB200.
• SFT: supervised fine-tuned on K-5 chat data, then one round of STaR rejection-sampling fine-tuning.
• RL (GRPO): segmented policy re-banding on a strictly K-5 verifiable-answer pool (Gemini-generated K-5 word problems + K-5-filtered GSM8K): 4 segments at rollout/training temperature 1.0 to a plateau, then temperature 2.0 (the bounded-corpus unlock temperature). TRL, fp32 master parameters.

Evaluation

MathCAMPS:

• K-5 pass@64 75.1 / pass@1 56.0
• beyond-K-5 pass@64 31.1 / pass@1 16.6

Usage

# transformers (chat)
from transformers import AutoModelForCausalLM, AutoTokenizer
repo = "manueldeprada/littlelearner-5b-bounded-grpo"
tok = AutoTokenizer.from_pretrained(repo)
model = AutoModelForCausalLM.from_pretrained(repo, dtype="bfloat16", device_map="cuda")
msgs = [{"role": "user", "content": "Liam has 3 apples and buys 4 more. How many apples does he have?"}]
ids = tok.apply_chat_template(msgs, add_generation_prompt=True, return_tensors="pt").to(model.device)
out = model.generate(ids)
print(tok.decode(out[0, ids.shape[1]:], skip_special_tokens=True))

# vLLM
from vllm import LLM
repo = "manueldeprada/littlelearner-5b-bounded-grpo"
llm = LLM(repo)
msgs = [{"role": "user", "content": "Liam has 3 apples and buys 4 more. How many apples does he have?"}]
print(llm.chat(msgs)[0].outputs[0].text)